I automate infrastructure and build AI tools: LLM inference on a single RTX 5090, agents with Claude Code, and the Proxmox and Docker stack they run on.
Public repositories on GitHub, most recently pushed first.
LLM server for one RTX 5090. Qwen3.8-27B at 91.0-93.7 tok/s (vLLM 68.9-70.0, llama.cpp 61.8-63.0). Runs Qwen3.8-Flash-Next (123.6 GiB) on a 32 GB card at 87.89-98.57 tok/s. Built for agents: tool calls, long context, many streams at once.
APA: adaptive-precision attention prefill for sm_120a (FP4 pass + exact FP16 hot tiles), MPL-2.0
Local-first AI prompt generator with a Gradio UI: 132 tuned roles for video, image, audio, 3D and creative AI tools. Runs uncensored Dolphin LLMs fully offline, auto-scaled to your GPU VRAM. No API keys, no cloud.
Rejection is the default: an operating manual for a Claude Code session whose job is driving other sessions and refusing anything not backed by evidence
Findet den schnellsten DNS-Resolver vom eigenen Standort aus. Pure stdlib-Python, laeuft im Container.
Recover the CUDA launch and implicit-destructor edges a syntactic indexer loses, written straight into CodeGraph's SQLite graph. On imp (kekzl/imp, 240k lines of C++/CUDA): kernel launch coverage 74.2% → 99.3%, destructors 0% → 85%. On llm.c, which hides nothing, it adds 0 edges.
Eine Honigbiene fliegt im Gegenlicht durch eine Sommerwiese — Echtzeit, WebGL2, ohne Build-Schritt
Monitor Entra ID (Azure AD) application secrets and certificates for expiration. Sends alerts via Email, Teams, Slack, or Webhook before credentials expire.
A living, continually learning neuromorphic being — a spiking neural network in C++23/CUDA with purely local plasticity (STDP, R-STDP, and Feedback Alignment instead of backprop). Runs on RTX 5090.
AI-powered email sorting for MailCow - Generate smart Sieve filters automatically using Claude API or local LLMs
Analyze meeting room usage via Microsoft Graph API with a React + FastAPI dashboard